使用CAMEL实现RAG过程记录

本文为学习使用CAMEL实现RAG的学习总结。

参考自官方cookbook，地址：https://docs.camel-ai.org/cookbooks/advanced_features/agents_with_rag.html

在官方cookbook分为了Customized RAG、Auto RAG、Single Agent with Auto RAG与Role-playing with Auto RAG四个部分。

Customized RAG

实现RAG需要有嵌入模型，为了简单验证，我这里使用的是硅基流动的嵌入模型。

现在先来看看在CAMEL中如何使用硅基流动的嵌入模型。

在.env文件中这样写：

Silicon_Model_ID="Qwen/Qwen2.5-72B-Instruct"
ZHIPU_Model_ID="THUDM/GLM-4-32B-0414"
DEEPSEEK_Model_ID="deepseek-ai/DeepSeek-V3"
Embedding_Model_ID="BAAI/bge-m3"
SiliconCloud_API_KEY="你的api key"
SiliconCloud_Base_URL="https://api.siliconflow.cn/v1"

我使用的是BAAI/bge-m3这个嵌入模型。

现在使用这个嵌入模型可以通过OpenAICompatibleEmbedding，因为它的格式与OpenAI是兼容的。

可以这样写：

from camel.embeddings import OpenAICompatibleEmbedding
from camel.storages import QdrantStorage
from camel.retrievers import VectorRetriever
import PyPDF2
import pathlib
import os
from dotenv import load_dotenv

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")

embedding_instance = OpenAICompatibleEmbedding(model_type=modeltype, api_key=api_key, url=base_url)

可以通过这个脚本快速测试是否能成功调用。

脚本可以这样写：

from camel.embeddings import OpenAICompatibleEmbedding

import pathlib
import os
from dotenv import load_dotenv

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")

embedding_instance = OpenAICompatibleEmbedding(model_type=modeltype, api_key=api_key, url=base_url)

# Embed the text
text_embeddings = embedding_instance.embed_list(["What is the capital of France?"])

print(len(text_embeddings[0]))

成功调用的输出如下所示：

有了嵌入模型之后，还需要有向量数据库。

这里使用的是QdrantStorage，要注意指定维度。

可以这样写：

storage_instance = QdrantStorage(
    vector_dim=1024,
    path="local_data",
    collection_name="test",
)

在CAMEL中实现Customized RAG需要使用VectorRetriever类，需要指定嵌入模型与向量数据库。

可以这样写：

vector_retriever = VectorRetriever(embedding_model=embedding_instance,
                                   storage=storage_instance)

在官方文档中使用的是一个PDF做的示例，但是我跑起来有问题，原因是一个相关依赖没有成功安装。

但是实现RAG很多时候确实是更加针对PDF文档的，如果不能适配PDF文档那就很鸡肋了。

因此可以自己简单写一下从pdf中提取文本的代码。

可以这样写：

import PyPDF2
def read_pdf_with_pypdf2(file_path):
    with open(file_path, "rb") as file:
        reader = PyPDF2.PdfReader(file)
        text = ""
        for page_num in range(len(reader.pages)):
            page = reader.pages[page_num]
            text += page.extract_text()
    return text
input_path = "local_data/test.pdf"
pdf_text = read_pdf_with_pypdf2(input_path)
print(pdf_text)

测试文档：

查看效果：

将这个代码写入脚本中，获取pdf文本之后，将文本进行向量化。

可以这样写：

def read_pdf_with_pypdf2(file_path):
    with open(file_path, "rb") as file:
        reader = PyPDF2.PdfReader(file)
        text = ""
        for page_num in range(len(reader.pages)):
            page = reader.pages[page_num]
            text += page.extract_text()
    return text

input_path = "local_data/test.pdf"
pdf_text = read_pdf_with_pypdf2(input_path)

print("向量化中...")

vector_retriever.process(
    content=pdf_text,
)

print("向量化完成")

然后可以进行提问，我这里问两个相关的与一个不相关的问题。

可以这样写：

retrieved_info = vector_retriever.query(
    query="你最喜欢的编程语言是什么？",
    top_k=1
)
print(retrieved_info)

retrieved_info2 = vector_retriever.query(
    query="你最喜欢的桌面开发框架是什么？",
    top_k=1
)
print(retrieved_info2)

retrieved_info_irrevelant = vector_retriever.query(
    query="今天晚上吃什么？",
    top_k=1,
)

print(retrieved_info_irrevelant)

效果如下所示：

通过以上步骤，我们就在CAMEL中实现了Customized RAG。

Auto RAG

现在再来看看如何在CAMEL中实现Auto RAG。

CAMEL中实现Auto RAG需要使用AutoRetriever类。

可以这样写：

from camel.embeddings import OpenAICompatibleEmbedding
from camel.storages import QdrantStorage
from camel.retrievers import AutoRetriever
from camel.types import StorageType
import pathlib
import os
from dotenv import load_dotenv

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")

embedding_instance = OpenAICompatibleEmbedding(model_type=modeltype, api_key=api_key, url=base_url)
embedding_instance.output_dim=1024

auto_retriever = AutoRetriever(
        vector_storage_local_path="local_data2/",
        storage_type=StorageType.QDRANT,
        embedding_model=embedding_instance)

retrieved_info = auto_retriever.run_vector_retriever(
    query="本届消博会共实现意向交易多少亿元？",
    contents=[
        "https://news.cctv.com/2025/04/17/ARTIbMtuugrC3uxmNKsQRyci250417.shtml?spm=C94212.PBZrLs0D62ld.EKoevbmLqVHC.156",  # example remote url
    ],
    top_k=1,
    return_detailed_info=True,
    similarity_threshold=0.5
)

print(retrieved_info)

可以直接传入一个链接。

查看效果：

Single Agent with Auto RAG

刚刚已经成功实现了相关信息的获取，但是还没有真正实现RAG。

实现的思路其实很简单，我这里提供了两个例子。

第一个将相关信息存入模型上下文：

可以这样写：

from camel.agents import ChatAgent
from camel.models import ModelFactory
from camel.types import ModelPlatformType,StorageType
from camel.embeddings import OpenAICompatibleEmbedding
from camel.storages import QdrantStorage
from camel.retrievers import AutoRetriever
import pathlib
import os
from dotenv import load_dotenv

sys_msg = 'You are a curious stone wondering about the universe.'

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Silicon_Model_ID")
embedding_modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")

siliconcloud_model = ModelFactory.create(
     model_platform=ModelPlatformType.OPENAI_COMPATIBLE_MODEL,
            model_type=modeltype,
            api_key=api_key,
            url=base_url,
            model_config_dict={"temperature": 0.4, "max_tokens": 4096},
)

embedding_instance = OpenAICompatibleEmbedding(model_type=embedding_modeltype, api_key=api_key, url=base_url)
embedding_instance.output_dim=1024

auto_retriever = AutoRetriever(
        vector_storage_local_path="local_data2/",
        storage_type=StorageType.QDRANT,
        embedding_model=embedding_instance)

# Define a user message
usr_msg = '3场官方供需对接会共签约多少个项目？'

retrieved_info = auto_retriever.run_vector_retriever(
    query=usr_msg,
    contents=[
        "https://news.cctv.com/2025/04/17/ARTIbMtuugrC3uxmNKsQRyci250417.shtml?spm=C94212.PBZrLs0D62ld.EKoevbmLqVHC.156",  # example remote url
    ],
    top_k=1,
    return_detailed_info=True,
    similarity_threshold=0.5
)

print(retrieved_info)

text_content = retrieved_info['Retrieved Context'][0]['text']
print(text_content)

agent = ChatAgent(
    system_message=sys_msg,
    model=siliconcloud_model,
)

print(agent.memory.get_context())

from camel.messages import BaseMessage

new_user_msg = BaseMessage.make_assistant_message(
    role_name="assistant",
    content=text_content,  # Use the content from the retrieved info
)

# Update the memory
agent.record_message(new_user_msg)

print(agent.memory.get_context())

# Sending the message to the agent
response = agent.step(usr_msg)

# Check the response (just for illustrative purpose)
print(response.msgs[0].content)

输出结果：

另一个就是通过写一个提示词，然后将获取的结果直接传给模型。

可以这样写：

from camel.agents import ChatAgent
from camel.retrievers import AutoRetriever
from camel.types import StorageType
from camel.embeddings import OpenAICompatibleEmbedding
from camel.models import ModelFactory
from camel.types import ModelPlatformType,StorageType
import pathlib
import os
from dotenv import load_dotenv

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Silicon_Model_ID")
embedding_modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")

siliconcloud_model = ModelFactory.create(
     model_platform=ModelPlatformType.OPENAI_COMPATIBLE_MODEL,
            model_type=modeltype,
            api_key=api_key,
            url=base_url,
            model_config_dict={"temperature": 0.4, "max_tokens": 4096},
)

embedding_instance = OpenAICompatibleEmbedding(model_type=embedding_modeltype, api_key=api_key, url=base_url)
embedding_instance.output_dim=1024

def single_agent(query: str) ->str :
    # Set agent role
    assistant_sys_msg = """You are a helpful assistant to answer question,
         I will give you the Original Query and Retrieved Context,
        answer the Original Query based on the Retrieved Context,
        if you can't answer the question just say I don't know."""

    # Add auto retriever
    auto_retriever = AutoRetriever(
            vector_storage_local_path="local_data2/",
            storage_type=StorageType.QDRANT,
            embedding_model=embedding_instance)

    retrieved_info = auto_retriever.run_vector_retriever(
        query=query,
        contents=[
        "https://news.cctv.com/2025/04/17/ARTIbMtuugrC3uxmNKsQRyci250417.shtml?spm=C94212.PBZrLs0D62ld.EKoevbmLqVHC.156",  # example remote url
    ],
    top_k=1,
    return_detailed_info=True,
    similarity_threshold=0.5
    )

    # Pass the retrieved information to agent
    user_msg = str(retrieved_info)

    agent = ChatAgent(
    system_message=assistant_sys_msg,
    model=siliconcloud_model,
    )

    # Get response
    assistant_response = agent.step(user_msg)
    return assistant_response.msg.content

print(single_agent("3场官方供需对接会共签约多少个项目？"))

输出结果：

Role-playing with Auto RAG

最后来玩一下Role-playing with Auto RAG。在我这里好像效果并不好，第二次调用函数就会出现问题。

可以这样写：

from camel.societies import RolePlaying
from camel.models import ModelFactory
from camel.types import ModelPlatformType, StorageType
from camel.toolkits import FunctionTool
from camel.embeddings import OpenAICompatibleEmbedding
from camel.types.agents import ToolCallingRecord
from camel.retrievers import AutoRetriever
from camel.utils import Constants
from typing import List, Union
from camel.utils import print_text_animated
from colorama import Fore
import pathlib
import os
from dotenv import load_dotenv

sys_msg = 'You are a curious stone wondering about the universe.'

base_dir = pathlib.Path(__file__).parent.parent
env_path = base_dir / ".env"
load_dotenv(dotenv_path=str(env_path))

modeltype = os.getenv("Silicon_Model_ID")
modeltype2= os.getenv("ZHIPU_Model_ID")
embedding_modeltype = os.getenv("Embedding_Model_ID")
api_key = os.getenv("SiliconCloud_API_KEY")
base_url = os.getenv("SiliconCloud_Base_URL")
siliconcloud_model = ModelFactory.create(
     model_platform=ModelPlatformType.OPENAI_COMPATIBLE_MODEL,
            model_type=modeltype,
            api_key=api_key,
            url=base_url,
            model_config_dict={"temperature": 0.4, "max_tokens": 4096},
)

siliconcloud_model2 = ModelFactory.create(
     model_platform=ModelPlatformType.OPENAI_COMPATIBLE_MODEL,
            model_type=modeltype2,
            api_key=api_key,
            url=base_url,
            model_config_dict={"temperature": 0.4, "max_tokens": 4096},
)

embedding_instance = OpenAICompatibleEmbedding(model_type=embedding_modeltype, api_key=api_key, url=base_url)
embedding_instance.output_dim=1024

auto_retriever = AutoRetriever(
        vector_storage_local_path="local_data2/",
        storage_type=StorageType.QDRANT,
        embedding_model=embedding_instance)

def information_retrieval(
        query: str,
        contents: Union[str, List[str]],
        top_k: int = Constants.DEFAULT_TOP_K_RESULTS,
        similarity_threshold: float = Constants.DEFAULT_SIMILARITY_THRESHOLD,
    ) -> str:
        r"""Retrieves information from a local vector storage based on the
        specified query. This function connects to a local vector storage
        system and retrieves relevant information by processing the input
        query. It is essential to use this function when the answer to a
        question requires external knowledge sources.

        Args:
            query (str): The question or query for which an answer is required.
            contents (Union[str, List[str]]): Local file paths, remote URLs or
                string contents.
            top_k (int, optional): The number of top results to return during
                retrieve. Must be a positive integer. Defaults to
                `DEFAULT_TOP_K_RESULTS`.
            similarity_threshold (float, optional): The similarity threshold
                for filtering results. Defaults to
                `DEFAULT_SIMILARITY_THRESHOLD`.

        Returns:
            str: The information retrieved in response to the query, aggregated
                and formatted as a string.

        Example:
            # Retrieve information about CAMEL AI.
            information_retrieval(query = "How to contribute to CAMEL AI?",
                                contents="https://github.com/camel-ai/camel/blob/master/CONTRIBUTING.md")
        """
        retrieved_info = auto_retriever.run_vector_retriever(
            query=query,
            contents=contents,
            top_k=top_k,
            similarity_threshold=similarity_threshold,
        )
        return str(retrieved_info)

RETRUEVE_FUNCS: list[FunctionTool] = [
    FunctionTool(func) for func in [information_retrieval]
]

def role_playing_with_rag(
    task_prompt,
    chat_turn_limit=5,
) -> None:
    task_prompt = task_prompt

    role_play_session = RolePlaying(
        assistant_role_name="Searcher",
        assistant_agent_kwargs=dict(
            model=siliconcloud_model,
            tools=[
                *RETRUEVE_FUNCS,
            ],
        ),
        user_role_name="Professor",
        user_agent_kwargs=dict(
            model=siliconcloud_model2,
        ),
        task_prompt=task_prompt,
        with_task_specify=False,
    )

    print(
        Fore.GREEN
        + f"AI Assistant sys message:\n{role_play_session.assistant_sys_msg}\n"
    )
    print(
        Fore.BLUE + f"AI User sys message:\n{role_play_session.user_sys_msg}\n"
    )

    print(Fore.YELLOW + f"Original task prompt:\n{task_prompt}\n")
    print(
        Fore.CYAN
        + "Specified task prompt:"
        + f"\n{role_play_session.specified_task_prompt}\n"
    )
    print(Fore.RED + f"Final task prompt:\n{role_play_session.task_prompt}\n")

    n = 0
    input_msg = role_play_session.init_chat()
    while n < chat_turn_limit:
        n += 1
        assistant_response, user_response = role_play_session.step(input_msg)

        if assistant_response.terminated:
            print(
                Fore.GREEN
                + (
                    "AI Assistant terminated. Reason: "
                    f"{assistant_response.info['termination_reasons']}."
                )
            )
            break
        if user_response.terminated:
            print(
                Fore.GREEN
                + (
                    "AI User terminated. "
                    f"Reason: {user_response.info['termination_reasons']}."
                )
            )
            break

        # Print output from the user
        print_text_animated(
            Fore.BLUE + f"AI User:\n\n{user_response.msg.content}\n"
        )

        # Print output from the assistant, including any function
        # execution information
        print_text_animated(Fore.GREEN + "AI Assistant:")
        tool_calls: List[ToolCallingRecord] = [
            ToolCallingRecord(**call.as_dict())
            for call in assistant_response.info['tool_calls']
        ]
        for func_record in tool_calls:
            print_text_animated(f"{func_record}")
        print_text_animated(f"{assistant_response.msg.content}\n")

        if "CAMEL_TASK_DONE" in user_response.msg.content:
            break

if __name__ == "__main__":
   role_playing_with_rag(
        task_prompt="4月17日凌晨，OpenAI正式宣布推出了什么？请参考https://www.thepaper.cn/newsDetail_forward_30670507",
        chat_turn_limit=5,
    )

调式运行：

会发现我们没有自己设置AI助手与AI用户的提示词，系统自动生成了提示词。

第一次成功进行了函数调用：

但是由于similarity_threshold设置的太高导致没有成功获取相关信息：

{'Original Query': '4月17日凌晨，OpenAI正式宣布推出了什么？', 'Retrieved Context': ['No suitable information retrieved from https://www.thepaper.cn/newsDetail_forward_30670507 with similarity_threshold = 0.75.']}

第二次调用把similarity_threshold调低了：

然后就会遇到错误：

暂时没有解决，不过也不影响学习Role-playing with Auto RAG。